Some of the most important technology in a phone is almost invisible to its owner. Memory does not have a camera lens or a bright screen. Yet every ambitious application depends on getting information to the processor quickly enough.
On 20 September 2026, Chinese memory manufacturer CXMT announced mass production of its G5 DRAM platform and two 24Gb LPDDR5X products. It is a manufacturing announcement with potential consequences for the components inside future devices. CXMT’s announcement, in Chinese
The interesting part is the manufacturing economics. A more capable memory supplier could give device makers another option. Whether that becomes a cheaper or better phone depends on what happens after the announcement.
One small letter changes the capacity claim
The new products are described as 24Gb, with a lowercase “b” for bits. That equals 3GB of raw capacity per die, using eight bits per byte. It does not mean CXMT has announced a single 24GB die, and it does not specify the total RAM of any particular phone.
CXMT also says G5 can produce at least 50% more dies per wafer than G4 when compared on an equivalent 8Gb basis. That qualification matters: the claim is about a normalised manufacturing comparison, not a measured 50% improvement in the yield of good chips. CXMT’s technical announcement and footnote
A wafer is the disc on which many chips are manufactured. Fitting more chip designs onto it can be valuable, but a factory also has to make those chips work reliably. Raw density and the proportion that pass testing answer different questions.
Why memory remains part of the AI story
LPDDR5X is a type of low-power working memory. Micron describes its uses across phones, notebooks and other systems, including applications that need rapid access to data for AI, imaging and gaming. It is not the same component as the storage that holds a photo library after a device is switched off. Micron’s LPDDR5X overview
Consider a phone processing an image while several applications remain open. A faster processor helps, but it still needs somewhere to hold and exchange the data it is using. Adding AI features makes it worth watching the memory subsystem alongside the headline processor.
There is also a capacity-versus-speed distinction. More memory can make room for larger workloads. Faster transfers can help feed those workloads. Neither improvement guarantees that an application will be well designed or that a handset will manage heat effectively.
A credible challenge still has to reach customers
The commercial test now is delivery. Our assessment is that repeat orders, qualification by device makers, reliable output and competitive costs will be more revealing than a single manufacturing milestone.
The announcement alone cannot establish a future market-share winner or a price cut at the checkout. Device pricing also includes many other components, software support, distribution and the manufacturer’s choices about margins.
Readers following the push to run more AI directly on personal computers will recognise the larger point: memory is part of the product, not an afterthought to the processor.
CXMT has supplied a fresh reason to watch that part of the industry. The next evidence should come from the devices and production volumes that follow.
Featured image: Representative circuit-board components, not an identified CXMT memory chip. Anne Nygård / Unsplash.
Leave a Reply